Individual Differences in Reading Skill and Language Lateralization: A Cluster Analysis

Abstract

Individual differences in reading and cerebral lateralization were investigated in 200 college students who completed reading assessments, divided visual field word recognition tasks, and received a structural MRI scan. Prior studies on this data set indicated that little variance in brain-behavior correlations could be attributed to the effects of sex and handedness variables (Chiarello, et al., 2009a,b; Welcome, et al., 2009). Here a more bottom-up approach to behavioral classification (cluster analysis) was used to explore individual differences that need not depend on a priori decisions about relevant subgroups. The cluster solution identified four subgroups of college age readers with differing reading skill and visual field lateralization profiles. These findings generalized to measures that were not included in the cluster analysis. Poorer reading skill was associated with somewhat reduced VF asymmetry, while average readers demonstrated exaggerated RVF/left hemisphere advantages. Skilled readers had either reduced asymmetries, or asymmetries that varied by task. The clusters did not differ by sex or handedness, suggesting that there are identifiable sources of variance among individuals that are not captured by these standard subject variables. All clusters had typical leftward asymmetry of the planum temporale. However, the size of areas in the posterior corpus callosum distinguished the two subgroups with high reading skill. Seventeen participants, identified as multivariate outliers, had unusual behavioral profiles and differed from the remainder of the sample in not having significant leftward asymmetry of the planum temporale. A less buffered type of neurodevelopment that is more open to the effects of random genetic and environmental influences may characterize such individuals.

It is widely acknowledged that individual differences moderate brain-behavior relationships (e.g., Colcombe, Kramer, Erickson, & Scalf, 2005; Ganis, Thompson, & Kosslyn, 2005). Left hemisphere language specialization, for example, varies in degree across individuals, and some persons demonstrate little asymmetry or reversed asymmetry, whether this is measured behaviorally, functionally, or anatomically (Chiarello, et al., 2009b; Kaiser, Kuenzli, Zappatore, & Nitsch, 2007; Prat, Long, & Baynes, 2007). There are both theoretical and applied reasons for investigating individual differences in brain organization. Theoretically, an emphasis on variation is essential in order to place psychological phenomena within a biological framework, where variability is of central interest rather than being treated as unfortunate “noise” (Kosslyn, et al., 2002). In applied settings, clinicians draw on our science to evaluate and treat individuals (Cherney & Small, 2006), yet they must frequently rely on knowledge of group averages to inform treatment strategies without a full appreciation of the range of normal (premorbid) structure-function associations.

With respect to reading, a highly skilled behavior that is unevenly developed in the population, research has emphasized comparisons between dyslexic and typical readers (e.g., Leonard & Eckert, 2008). However, the typical reading controls in such studies are usually treated as an undifferentiated group and provide little information about the relationships between variations in reading skill and lateralization within the normal reading population. In the current study, we explore a multivariate approach (cluster analysis) to investigate whether differing behavioral/lateralization profiles can be observed among typical college age readers.

Prior neuropsychological studies of individual differences have employed one of two methods. Some insight can be derived from regression approaches, in which one examines whether the magnitude of a neural variable predicts the degree of a behavioral variation. For example, Forstman, et al. (2008) reported that activation of the right inferior frontal cortex correlated with a reaction time measure of response inhibition. In some cases multiple regression techniques can be used to determine the influence of additional variables on brain-behavior relationships and/or to hold constant potentially confounding variables (e.g., Peterson, Gable, & Harmon-Jones, 2007). This approach can capture broad patterns across the studied population, but leaves open the question of whether there may be meaningful subpopulations with different structure/function relationships. For example, if one observes a .4 correlation, are the individuals with scores near the regression line different in important ways from those with scores that are more distant from this line? Perhaps the latter individuals have behavior that is predicted by a different set of neurofunctional variables. This idea is difficult to assess directly within a regression framework.

An alternate approach is to identify groups a priori and then investigate whether these groups differ in behavioral or neuropsychological variables. The voluminous neuropsychological literature on differences attributable to sex, handedness, or diagnostic category attests to the popularity of this approach (e.g., Harrington & Tomaszewski, 2008; Narr, et al., 2007; Sommer, et al., 2004, 2008; Wallentin, 2009). This method has the advantage that groups can be objectively and reliably determined, and it addresses the natural curiosity about potential neural differences among individuals who can be readily distinguished from each other. We have previously employed this approach to investigate sex differences (Chiarello, et al., 2009b; Leonard, et al., 2008; Leonard, Towler, Welcome, & Chiarello, 2009), handedness differences (Chiarello, et al., 2009a), and sex/handedness groups (Welcome, et al., 2009) in the sample reported here. However, very little variance in brain structure and behavior was explained by these variables. No sex differences in corpus callosum area or in the volumes of language-relevant cortex were obtained after controlling for brain volume (Leonard, et al., 2008; Welcome, et al., 2009), nor were there sex differences in perisylvian asymmetries (Chiarello, et al., 2009b). With respect to behavioral (divided visual field) asymmetries, very small sex differences (accounting for 2% of the total variance) were observed in only 2 of 8 lexical tasks and these could not be replicated in independent subsamples (Chiarello, et al., 2009b). Similarly, there were no significant differences between consistent handers (those with strong hand preference for 5 of 5 activities) and mixed handers¹ in VF asymmetry (Chiarello, et al., 2009a), corpus callosum area (Welcome, et al., 2009), or perisylvian asymmetries (unpublished data). Hence, the current investigation was motivated by the realization that most of the individual variation in reading skill, and in cerebral asymmetry, was not captured by categorizing people into sex or handedness groups.

Some investigations have gone beyond these ubiquitous grouping variables to explore differences related to particular individual traits. For example, Golestani, et al. (2007) found that fast learners of nonnative speech sounds (N = 11) had greater white matter density in left Heschl’s gyrus, relative to slow learners (N =10). Such studies can yield important insights, but must rely on the investigators’ a priori ideas to identify subject characteristics that may most meaningfully reveal individual differences in brain function. Because individuals differ from each other along multiple dimensions, forming groups based on a single trait can only reveal a fraction of the diversity in the population.

Data gathered via the Biological Substrates for Language Project enabled us to explore an alternate approach to understanding individual differences. This approach (cluster analysis) makes no a priori assumptions about the relevant dimensions along which individuals might differ, in contrast to typical neuropsychological analyses of individual differences. The project collected demographic and reading test data, and divided visual field (VF) asymmetries for a variety of lexical tasks from 200 college students who also received a structural MRI scan. One goal of the project is to document the range of anatomical and behavioral asymmetries from a population of normal readers, and to begin to explore the relationship between neuroanatomical and behavioral variation. In the current investigation we made no assumptions about the relevant dimensions along which individuals might differ behaviorally. Instead we employed a statistical technique (cluster analysis) that groups similar individuals together based on their scores on multiple variables. This bottom-up approach is exploratory rather than confirmatory in that it allows groups to emerge from the patterns inherent in the dataset, rather than testing whether the data conform to groups determined a priori.

After clusters were identified from the reading and visual field scores, we examined whether the groups so formed differed in neuroanatomical regions that may support lateralized functioning (planum temporale asymmetry and corpus callosum area). Leftward asymmetry of the planum temporale is quite robust, and encompasses posterior language-relevant cortex (Shapleske, Rossell, Woodruff, & David, 1999), and visual field asymmetries may vary with corpus callosum structure (e.g., Hellige, Taylor, Lesmes, & Peterson, 1998; Zaidel & Iacaboni, 2003). The current study was designed to explore the following questions:

1. Are there subgroups within the college population characterized by differing profiles of reading and VF/hemisphere asymmetry? If so, to what extent do these groups differ in standard demographic variables such as sex and handedness? The answers to these questions can shed light on the relationship between reading skill and degree of visual lateralization.

2. Do the groups identified by the cluster analysis differ in either planum temporale asymmetry or corpus callosum area? If so, this could suggest relationships between behavioral and neuroanatomical profiles that might not be transparent from alternate approaches that rely on sample-wide correlations or a priori subject groupings.

3. Are there some individuals whose reading behavior and VF asymmetry depart from typically observed outcomes? In the current study these would be persons who showed the greatest multivariate deviation from the sample dataset (i.e., outliers). Do such individuals show unusual neuroanatomical features? We have suggested previously that some individuals may have less typical or well-regulated trajectories of neural development that can result in unusual neuroanatomical and behavioral asymmetries (Chiarello, Kacinik, Manowitz, Otto, & Leonard, 2004). Hence, we would predict a higher likelihood of unusual neuroanatomical features in the MRI scans obtained from those identified behaviorally as multivariate outliers.

Method

Participants took part in five sessions of behavioral testing, and then received a structural MRI scan in their final session. Behavioral testing and preliminary analyses of these data were conducted at the University of California, Riverside with the experimenters blind to the status of the brain measurements. Similarly, brain measurements were made at the University of Florida by anatomists who were blind to the identity and behavioral findings of the participants. The behavioral and anatomical data were pooled only after the data were scored and brain measurements completed.

Participants

Campus-wide electronic messages and announcements were used to recruit potential participants. In order to obtain a representative sample, we did not attempt to recruit equal numbers of left- and right-handers. The study was approved by Institutional Review Boards at both the University of California and the University of Florida. One hundred male and 100 female university student volunteers participated, receiving $100 payment for their participation. Subjects with a history of brain injury or disease or conditions incompatible with an MRI scan were excluded. A neuroradiologist reviewed all scans for pathology, and four additional participants were excluded from the final sample due to abnormal findings on the MRI. All participants were native speakers of English with normal or corrected-to-normal vision. To assess handedness, a five-item preference questionnaire was utilized (Bryden, 1982), which yields an index ranging from +1.00 (extreme right handedness) to -1.00 (extreme left-handedness). We also classified participants as either consistent- or mixed-handers (Chiarello, et al., 2009a) based on their questionnaire responses; 103 reported no use of the nondominant hand (consistent) and 97 reported at least some use of the nondominant hand (mixed). More detailed demographic data on this sample is described in prior publications (Chiarello, et al., 2099a,b; Leonard, et al., 2008; Welcome, et al., 2009), where it was noted that a range of reading skill was observed.

Behavioral Stimuli and Procedure

In an initial 2-hour session participants completed questionnaires regarding handedness, language and family background, reading history [Adult Reading History Questionnaire (AHRQ), Lefly & Pennington, 2000], and standardized measures of reading skill and intelligence (Wechsler, 1999; Woodcock, 1998). Following this session, four test sessions were held on separate days in which participants completed 7 lateralized word recognition tasks. All participants received tasks and test sessions in the same order.

Experimental stimuli consisted of 3–6 letter concrete nouns and/or pronounceable nonwords. Nonwords were created by replacing a single letter of a concrete noun, with each position of replacement occurring equally often. No stimuli were repeated within an experimental session, and no stimulus was used more than twice throughout the entire study. Word lists for each task were equated for word length and log-transformed word frequency based on the Hyperspace Analogue to Language corpus (Lund & Burgess, 1996). Mean word length for each task ranged from 4.44 to 4.64 and mean log word frequency ranged from 4.16 to 4.71. Within each task, items were matched across visual field conditions on the basis of length, log frequency (Lund & Burgess, 1996), familiarity (Wilson, 1988) and imageability (Wilson, 1988).

All stimuli were presented in uppercase, black 20 point Helvetica font on a white background on an Apple Studio Display M7649 monitor. Macintosh computers were used for stimulus presentation and recording of manual responses in the visual field tasks. Psyscope programming software (Cohen, MacWhinney, Flatt, & Provost, 1993) was used to control experimental events and record responses. Participants were seated 60 cm in front of the monitor, using a headrest to stabilize head position. For those experiments requiring manual responses (Lexical Decision, Masked Word Recognition, and Semantic Decision), participants used the index fingers of each hand on the ‘.’ and ‘x’ keys to indicate one response and the middle fingers of each hand on the ‘/’ and ‘z’ keys to indicate the other response. This configuration was designed to accommodate both left- and right-handed participants. A Sony ECM-MS907 microphone was used to register vocal responses. Vocal responses were entered into the data file by an experimenter. Special codes were entered for spurious vocal responses (a cough, for example), or failure to respond, and such trials were not analyzed.

The tasks were administered across 4 testing sessions, in the following order:

Lexical Decision: 90 word and 90 nonword trials, keypress discrimination response, 125 ms exposure.

Word Naming: 90 trials, pronounce word, 125 ms exposure.

Category Generation: 82 trials, produce exemplar of stimulus noun category (e.g., FRUIT), 155 ms exposure.

Nonword Naming: 90 trials, pronounce nonword, 150 ms exposure.

Masked Word Recognition: 100 trials, recognize word preceded and followed by 60 ms pattern mask (@#@#), two-alternative forced choice key press response, 30 ms exposure. The response alternatives differed by only a single letter.

Verb Generation: 100 trials, produce verb associated with stimulus noun, 150 ms exposure.

Semantic Decision: 120 trials, determine whether stimulus noun represents a naturally occurring or manmade object, keypress response, 120 ms exposure.

On average, each session was separated by 4 days. Each task was preceded by 30–48 practice trials.

Stimuli were randomly presented to the left or right visual field (LVF, RVF), 1.8 degrees eccentric from a central fixation “+”. At the onset of each trial, the fixation marker appeared for 600–805 ms and flickered just prior to the onset of the stimulus. Participants were instructed to maintain central fixation and respond as quickly and accurately as possible.

Imaging Procedure

After the images were reviewed for neuropathology they were transferred to compact discs at the Imaging Center and sent to the McKnight Brain Institute at the University of Florida. Preprocessing the images was performed using FSL scripts (http://www.fmrib.ox.ac.uk/) (Smith et al., 2004). Extraction of the brain parenchyma from scalp and skull was performed with BET (Smith, 2002) before registration (FLIRT) (Jenkinson and Smith, 2001) to a 1 mm isovoxel study-specific template image aligned into the Talairach planes. No nonlinear warping was performed on the images. Hence, changes in the images were restricted to the translation and rotation necessary to align the midline and the anterior commissure-posterior commissure axis with the standard Talairach planes. Segmentation into separate grey matter, white matter and cerebrospinal fluid (CSF) volumes was performed using FAST (Zhang, Brady & Smith, 2001). In these volumes, each voxel is represented as a partial volume estimate of a particular tissue type. The volume of each tissue type was calculated by multiplying the number of voxels times the average partial volume estimate of those voxels as described on the FSL website. Volumes, surface areas, means, standard deviations, and average asymmetries were automatically accumulated in a data file for statistical analysis. Each structure was measured twice by at least two different investigators who were blind to hemisphere and subject characteristics. When there was more than 15% disagreement between the average values for the two measurements, the experimenters conferred and identified the reason for disagreement and then remeasured until the two measures agreed.

Grey, white and cerebrospinal fluid (CSF) volumes of each hemisphere were estimated by outlining every fifth sagittal image starting at the midline. The brainstem was excluded by transection in the midcollicular plane. The midsection was traced twice and half the slab volume added to each hemisphere. The inter-rater reliability of this measure is > .98 (intraclass correlation). Preliminary studies showed that the accuracy of volumes sampled in this way was equivalent to that in which every section was measured.

Surface area of the planum temporale was calculated between × = 47 and 56 (sagittal coordinates adjusted for hemisphere width and chosen to maximize lateral asymmetry as well as reliability) (Chiarello, et al., 2004; Eckert, et al., 2001; Leonard, et al., 1996). In individuals with one clearly defined Heschl’s gyrus, the anterior border of the planum temporale was defined as the depth of the sulcus that formed the posterior border of Heschl’s gyrus (Heschl’s sulcus). When Heschl’s gyrus is indented by an intermediate sulcus, the tracing includes the gyri on both banks of the sulcus. When an independent gyrus appears posterior to Heschl’s gyrus, this gyrus is included in the planum measurement (Eckert, et al., 2006). The posterior boundary of the planum temporale was defined as the origin of the posterior ascending ramus or the termination of the Sylvian fissure. Inter-rater reliability for these measurements is .85. Asymmetry coefficients for the planum temporale were calculated by subtracting the left measure from the right and dividing by the average, so that leftward asymmetries yielded positive coefficients. We observed reliable leftward asymmetry for the planum temporale in the current sample (see Chiarello, et al., 2009b; Leonard, et al., 2009). A comparative study of techniques to measure the planum temporale (Best and Demb, 1999) found that asymmetry measures using this index agreed well with those gained using other techniques.

The area of the corpus callosum was extracted from the midsagittal white matter image. It was subdivided into seven subdivisions (rostrum, genu, anterior, mid and posterior body, isthmus and splenium) using the method of Witelson (1989). Because of the well-known relationship between corpus callosum area and overall cerebral volume, we controlled for the effects of cerebral volume through regression (Smith, 2005). Hence we report the residualized callosal area, which statistically eliminates the variance in callosal area that is due to brain size (see Welcome, et al., 2009 for prior use of this measure in the current sample).

RESULTS

Data Analytic Approach

Cluster analysis is an exploratory technique that identifies empirically-determined groups of individuals who are similar to each other on a particular set of variables. This technique maximizes within-cluster homogeneity and between-cluster heterogeneity (Hair & Black, 1998). We used Ward’s Method, a hierarchical agglomerative procedure that identifies clusters in which the variance of cases within a cluster is relatively small. The distance metric is the sum of the squared distances of each individual’s data from the mean of its cluster. This method is widely used in social science research (Romesburg, 1984).

Because variables that are strongly intercorrelated can be weighted more in cluster analyses, it is preferable to select measures that are not strongly correlated (Everitt, 1974; Hair & Black, 1998). We examined the intercorrelation matrices for our reading and lateralization tasks, and selected as cluster variates those that were uncorrelated or only very weakly correlated. These represented both reading skill and DVF performance measures (both RT and accuracy) (average intercorrelation of the cluster variates = .076). Word Attack scores showed the largest range of individual variation of the reading subtests, so this reading measure was included. The cluster variate asymmetry scores included both word identification and semantic tasks: masked word recognition (RT and accuracy), verb generation (RT and accuracy), lexical decision (RT and accuracy), and nonword naming RT. Based on our previous reliability analyses of various asymmetry indices (Chiarello, et al., 2009b), accuracy asymmetry was estimated by the lambda z-score (Bryden & Sprott, 1981) and RT asymmetry by the laterality index (LVF-RVF)/(LVF+RVF). To eliminate scaling differences across our measures, all data were z-scored before being entered into the cluster analysis (Hair & Black, 1998).

We first identified outliers from the multivariate dataset that included only the cluster variates, and then ran and examined the cluster analysis with outliers excluded. We next compared the clusters on reading/lateralization task variables that were not used to create the clusters to examine the generalizability of the classification. Planar asymmetry and corpus callosum measurements were then compared across clusters to investigate potential neural correlates of the behavioral profiles. Although the outliers do not represent their own cluster, we also examined their data since it was predicted that individuals with atypical combinations of behavioral features might also have unusual reading or neuroanatomical profiles.

Findings

Multivariate outliers were identified from the set of cluster variates using the SAS OUTLIER macro (Friendly, 2003). The procedure calculates the robust Mahalnobis distances for each case in the data set, and identifies the probability that a given case belongs to any identifiable cluster. Seventeen cases were identified as outliers (probability < .05 for cluster membership, d² values of 15.5 or above), and they were excluded from the cluster analysis. The SAS CLUSTER procedure (Ward’s method) was used to identify cluster solutions. As there was a large drop in the eigenvalues between the 5- and 6-cluster solutions (from .93 to .77), we examined the 3-, 4-, and 5-cluster solutions.

The 3-cluster solution (eigenvalue = 1.13) separated individuals into groups with low, high, or average word attack scores. On nearly all asymmetry measures, those with average word attack scores had larger RVF advantages than those with poor word attack scores. However, the asymmetries for those with high word attack scores did not show any discernable pattern relative to the other two groups. However, in the 4-cluster solution (eigenvalue = .97), the high word attack individuals split into two clusters, one with much smaller asymmetries than the others in all but one task measure. In the 5-cluster solution (eigenvalue = .93), those with average word attack scores split into two clusters; although these two new clusters differed significantly in asymmetry for several tasks, the direction of the differences was inconsistent (i.e., one cluster had larger RVF advantages for 4 task measures, and the other had larger RVF advantages for 3 task measures).

Based on these findings the 4-cluster solution appeared to be the most meaningful, accounting for 74% of the variance. Hence we adopted this solution. Table 1 displays the mean z-scores² for the cluster variates for each cluster, and the univariate F-ratios that tested cluster differences for each variable. The F ratios indicated that all variables used in the cluster analysis successfully differentiated the clusters, verifying that none was a masking variable (Everitt, Landau, & Leese, 2001). Cluster 1 (N = 61) was characterized by very low word attack scores (for this sample), and VF asymmetries that were mostly smaller than average. The individuals in Cluster 2 (N = 26) had relatively high word attack scores and VF asymmetries that were uniformly near bottom of distribution, including some that were extremely low (masked word recognition and lexical decision accuracy, verb generation RT)³. Cluster 3 (N = 63) was characterized by average word attack scores, and, with the exception of masked word recognition, VF asymmetries that were quite large. Cluster 4 resembled Cluster 2 in having high word attack scores, but VF asymmetries that varied substantially over tasks – individuals in this cluster had the highest asymmetries in masked word recognition, average asymmetries in lexical decision, and reduced asymmetries for verb generation accuracy and nonword naming RT. For Cluster 4 the standard deviation for the cluster variate asymmetries was .495 (deviations for the other clusters were between .206 and .320).

Table 1.

Cluster Z-score Means (sd) and Univariate F Statistics for Cluster Variates

	Cluster 1 (N=61)	Cluster 2 (N=26)	Cluster 3 (N=63)	Cluster 4 (N=33)	F (3,179)	η2
Reading Skill VF Asymmetry	Poorer Low-to- Average	Good Low	Average Large	Good Varies by Task
Reading Subtest
Word Attack	−.748 (.634)	.851 (.547)	−.032 (.975)	.600 (.794)	34.9***	.37
Accuracy Asymmetry
Masked Word Rec	.135 (.890)	−.715 (.836)	−.195 (.855)	.672 (.654)	15.3***	.20
Verb Generation	−.112 (.920)	−.164 (.659)	.358 (1.08)	−.430 (.792)	6.04***	.09
Lexical Decision	−.295 (.869)	−.777 (.731)	.505 (1.02)	.006 (.719)	15.7***	.21
RT Asymmetry
Nonword Naming	−.106 (.607)	−.682 (.684)	.328 (.716)	−.279 (.734)	15.3***	.20
Masked Word Rec	−.356 (.608)	−.146 (.870)	−.197 (.912)	.937 (1.00)	18.8***	.24
Verb Generation	−.250 (.909)	−.703 (.898)	.338 (.942)	.253 (.978)	9.9***	.14
Lexical Decision	−.502 (.769)	−.379 (.730)	.607 (.745)	−.017 (.817)	24.1***	.29

^***p < .0001

Table 2 presents the z-scored means, and univariate F tests, for the additional reading and VF lateralization measures that were not used in the cluster analysis. With the exception of the RT asymmetry for category generation, all of these measures differed by cluster. This indicates that the group differences identified in the cluster analysis generalize to additional reading and VF lexical tasks. Examination of the mean scores for each cluster confirms our original characterization of the clusters. Cluster 1 individuals, who represented approximately 30% of the sample, performed most poorly on the other reading subtests and had low-to-average VF asymmetries. Cluster 2 individuals performed well on the reading measures and obtained consistenly low VF asymmetries. Cluster 3 (approximately 32% of the sample) was characterized by near average reading skills and a tendency toward large VF asymmetries. Cluster 4 individuals demonstrated excellent reading skill, and variable VF asymmetries (although the differences across tasks were not as dramatic as for the cluster variates).

Table 2.

Cluster Z-score Means (sd) and Univariate F Statistics for Non-cluster Variates

	Cluster 1 (N=61)	Cluster 2 (N=26)	Cluster 3 (N=63)	Cluster 4 (N=33)	F(3,179)	η2
Reading Skill VF Asymmetry	Poorer Low-to- Average	Good Low	Average Large	Good Varies by Task
Reading Subtest
Word Identification	−.535 (.857)	.434 (.884)	.037 (.971)	.528 (.913)	12.7***	.19
Passage Compreh	−.382 (.912)	.376 (.867)	−.117 (1.07)	.362 (.882)	6.3**	.11
Accuracy Asymmetry
Word Naming	−.047 (1.00)	−.635 (.911)	.398 (.902)	−.321 (1.01)	8.7***	.13
Categ Generation	−.048 (1.00)	−.280 (.881)	.316 (1.11)	−.325 (.808)	4.0*	.06
Nonword Naming	−.130 (.801)	−.328 (.936)	.434 (1.00)	−.270 (.889)	7.3***	.11
Semantic Decision	.247 (.891)	−.752 (.950)	.110 (1.02)	−.111 (.963)	7.1**	.11
RT Asymmetry
Word Naming	−.335 (1.06)	−.608 (.795)	.428 (.931)	.068 (.820)	10.5***	.14
Categ Generation	.126 (1.08)	−.188 (.827)	−.074 (1.11)	.044 (.781)	< 1	.01
Semantic Decision	−.008 (.823)	−.434 (1.07)	.091 (.977)	.316 (1.00)	3.2*	.05

^***p < .0001,

^**p < .001,

^*p < .05

It is notable that not every possible combination of reading skill and VF lateralization was observed. For example, there was no cluster of individuals with highly skilled reading and consistently large RVF/left hemisphere advantages, or with average reading ability and reduced VF asymmetry. Although, as discussed further below, there is no simple linear relationship between reading skill and visual lateralization, this does not indicate that every possible relationship between these domains is observed.

We also examined whether the clusters could be differentiated by demographic variables (see Table 3). Chi-square analyses examined whether the categorical variables sex and mixed vs consistent handedness differed by cluster. Neither differed significantly by cluster (sex χ²(3,N=183) = 5.51, p = .14; mixed/consistent handedness χ²(3,N=183) = 2.94, p =.40). A more continuous measure of handedness, score on the inventory, also did not differ by cluster (F < 1). The clusters did not differ in socioeconomic status (F < 1), or self-reported history of reading problems as indexed by the ARHQ (F (3,177) = 1.48, p = .22). However, there were some differences in mean age, F(3,179) = 5.02, p < .01, η² = .08, verbal IQ, F(3,179) = 5.72, p < .001, η = .09, and performance IQ, F(3,179) = 4.56, p < .01, η = .07. Post-hoc contrasts (Tukey-Kramer adjustment) indicated that Cluster 1 individuals were younger (M = 20.4 yrs) than those in Clusters 2 (M = 23.3 yrs), p = .002, d= .86 and 3 (M = 22.0 yrs), p = .047, d = .48. Those in Cluster 1 also had lower verbal IQs (M=104.5) than those in Clusters 2 (M = 112.3), p = .01, d = .75, and 4 (M = 112.7), p = .003, d = .77. Performance IQ was lower for Cluster 1 (M=105.2) relative to Cluster 2 (M = 113.5), p = .01, d = .76. None of the other contrasts was significant.

Table 3.

Demographic Characteristics of Clusters and Unclassifiable Individuals (Outliers)

	Cluster 1 (N=61)	Cluster 2 (N=26)	Cluster 3 (N=63)	Cluster 4 (N=33)	Outliers (N = 17)
Mean Age (yrs.)	20.4	23.3	22.0	21.6	22.1
Mean Hand Preference Score	+.677	+.696	+.725	+.661	+0.812
% Mixed (Left) Handed	45.9% (13.1%)	61.5% (11.5%)	46.0% (11.1%)	57.6% (9.1%)	29.4% (5.9%)
% Male	41.1%	61.5%	60.3%	45.5%	41.1%
Mean SES	3.16	3.37	3.47	3.33	3.29
VIQ	104.5	112.3	108.0	112.7	113.4
PIQ	105.2	113.5	107.2	111.6	114.8
Mean ARHQ	.316	.289	.292	.272	.302

To recap our behavioral findings, the cluster analysis identified four subtypes of individuals. One type (Cluster 1) showed poor performance on the reading and IQ measures and had small VF asymmetries. A second type (Cluster 3) showed average standardized test performance and large VF asymmetries. Two smaller sub-groups of individuals showed good reading performance, and this was associated with either with low VF asymmetries (Cluster 2) or with variable task asymmetries (Cluster 4).

There were no differences between clusters in brain volume, F(3,179) = 1.30, p = .28. Table 4 provides the means for each cluster for planum temporale asymmetry and the residualized corpus callosum measurements (in z-scores). Planum temporale asymmetry did not differ significantly by cluster. We did observe cluster differences for some of the callosal measurements (see Table 4). Although there were no differences for the entire corpus callosum area, or for the more anterior subregions, differences were observed for the posterior body and splenium. Post hoc contrasts indicated that the posterior body was significantly smaller for Cluster 4 individuals (good readers, task-dependent VF asymmetries), than for Cluster 1 individuals (poor readers, low/average VF asymmetries), p = .004, d = .76. Although it did not reach significance, a similar pattern was observed for the callosum midbody. Post-hoc contrasts for the splenium indicated that Cluster 2 individuals (good readers, reduced VF asymmetries) had larger splenia relative to Cluster 1 (p = .02, d = .61), Cluster 3 (p = .03, d = .73), and Cluster 4 (p = .06, d = .68)⁴. Because Cluster 2 had the smallest number of cases, one might suspect that a few extreme scores in this group could unduly skew the mean. However, as Table 4 indicates, Cluster 2 actually had the smallest standard deviation in splenium size, rendering such an account implausible.

Table 4.

Mean (sd) Planum Temporale (PT) Asymmetry and Residualized Corpus Callosum Measurements (z-scores) by Cluster

	Cluster 1 (N=61)	Cluster 2 (N=26)	Cluster 3 (N=63)	Cluster 4 (N=33)	F(3,179)	η2
Reading Skill VF Asymmetry	Poorer Low-to- Average	Good Low	Average Large	Good Varies by Task
PT Asymmetry	.002 (.887)	.164 (1.04)	−.046 (1.12)	.180 (.984)	< 1
Total Callosal Area	.062 (1.12)	.153 (1.00)	−.055 (.948)	−.212 (.932)	< 1
Rostrum	.028 (.883)	−.093 (1.04)	−.027 (1.10)	−.075 (1.07)	< 1
Genu	.097 (1.05)	−.165 (1.10)	.001 (.927)	−.078 (1.02)	< 1
Anterior Body	−.030 (1.07)	.070 (1.04)	.066 (1.00)	−.051 (.955)	< 1
Midbody	.191 (1.09)	.124 (.898)	−.041 (.946)	−.336 (.938)	2.19, p = .09	.04
Posterior Body	.224 (1.03)	−.059 (.940)	−.029 (.940)	−.478 (.789)	3.92**	.06
Isthmus	−.052 (1.19)	.097 (1.16)	−.016 (.847)	−.095 (.808)	< 1
Splenium	−.068 (1.05)	.537 (.935)	−.146 (.937)	−.117 (.980)	3.28*	.05

^**p < .01,

^*p < .05

As noted earlier, 17 individuals were identified as multivariate outliers and could not be included in the cluster analyses. These outliers, then, have reading/VF asymmetry profiles on the cluster variates that are not characteristic of the rest of the sample and are not similar enough to each other to form another cluster. However, we examined their behavioral and neuroanatomical data to help understand why these individuals might be unusual and whether they might have atypical neuroanatomical features. Table 5 presents the means and standard deviations for the outliers for the reading subtests and for composite measures of overall VF asymmetry and consistency of asymmetry. The composite asymmetry score for each participant is the average of their z-scored asymmetries over all VF tasks, and the consistency of asymmetry is each participant’s standard deviation of their asymmetry across measures (see Chiarello, et al., 2004 see Chiarello, et al., 2009b for previous use of these measures). The latter measure of variation may be particularly important when examining atypical individuals, if their phenotype is more open to the effects of random variation. Because we wished to determine ways in which the outliers differed from the rest of the sample, t-tests were computed to make this comparison. In cases of unequal variances, adjusted t’s and degrees of freedom were used.

Table 5.

Comparison of Clustered Subjects and Outlier Means (z-scores) for Behavioral Measures

	Clustered Subjects (N = 183)	Outliers (N = 17)	t	d
Reading Subtest
Word Attack	−.031 (.982)	.337 (1.16)	1.45, ns
Word Identification	−.009 (.996)	.092 (1.07)	< 1
Passage Comprehension	−.049 (.999)	.526 (.878)	2.29*	.61
Composite VF Asymmetry
Accuracy	−.011 (.536)	.124 (.670)	< 1
Response Time	−.023 (.451)	.249 (.504)	2.36*	.57
Consistency of VF Asymmetry
Accuracy	.851 (.251)	.999 (.326)	2.26*	.51
Response Time	.851 (.277)	1.43 (.315)	8.15****	1.95

^*p < .05;

^****p < .00001

With respect to the reading measures, outliers had unusually high passage comprehension scores, but not word-level reading scores (see Table 5). Their composite RT asymmetry provided evidence of larger RVF/LH advantages, relative to the rest of the sample, when combined across tasks. It is interesting to note that the outliers had much higher within-subject variability in their task asymmetries, especially for response time (see Table 5, consistency of VF asymmetry). This indicates that, not only do the outliers fail to resemble identifiable subgroups in their reading/asymmetry profiles, but their individual asymmetries vary more from task to task than do the rest of the sample – they show more extreme variation in VF asymmetry across tasks than was observed for the rest of the sample.

For contrasts involving demographic variables, the clustered and outlier subjects did not differ by sex, age, handedness, socioeconomic status or history of reading problems. However, the outliers had somewhat higher IQs, relative to the clustered individuals (VIQ, 113.4 vs 108.3, t (198) = 1.85, p = .07, d = .48; PIQ, 114.8 vs 108.3, t(198) = 2.27, p < .05, d = .64).

There was no difference in brain volume between the outliers and clustered individuals (t < 1). The outliers were less likely to have leftward asymmetry of the planum temporale, t(198) = 1.89, p = .059, d = .51. As a group, their mean asymmetry did not differ from zero, t(16) = 1.34, p =.20. In comparison, the clustered individuals as a whole, and every separate cluster, showed leftward planar asymmetries that significantly differed from zero (Cluster 1: t(60) = 6.31, p < .0001; Cluster 2 t(25) = 4.32, p < .0002; Cluster 3: t(62) = 4.75, p <.0001; Cluster 4: t(32) = 5.22, p < .0001). Inspection of the untransformed asymmetries indicated that 41.2% of the outliers had larger right, than left, PT (compared to 19.7% for the individuals who were successfully assigned to one of the four clusters) - see Figure 1. When corrected for hemispheric gray matter volume, the outliers’ left planum did not differ from that of the rest of the sample (t < 1), while their right planum was significantly larger, t(198) = 2.14, p < .05, d = .30. Although several callosal subregions were larger for the outliers (rostrum, posterior body, isthmus) these differences did not reach significance.

Figure 1.

Distribution of planum temporale asymmetry (untransformed) for multivariate outliers vs clustered subjects. Positive values indicate leftward asymmetry.

DISCUSSION

To summarize our major findings, the cluster analysis identified four sub-groups with differing reading/VF lateralization profiles. The clusters did not differ by sex or handedness, suggesting that there is identifiable behavioral variance between individuals that is not captured by standard subject variables. Although these groups has similar planar asymmetries, some group differences were observed in posterior callosal areas, as discussed further below. A small portion of the sample was identified as multivariate outliers with respect to the variables used in the cluster analysis. These individuals were notable for their high passage comprehension scores and highly variable VF asymmetries. Interestingly they also had an unusual distribution of planum temporale asymmetry. These results suggest that alternate quantitative approaches to the investigation of individual differences in brain-behavior relations have promise.

We begin by interpreting the differences between the four clusters and then discuss the ways in which the multivariate outliers differ from those with more predictable reading/VF asymmetry profiles. After considering the strengths and limitations of this study, we conclude by suggesting a novel framework for interpreting individual differences in brain-behavior relations.

Interpretation of Clusters

In characterizing the differences between the clusters, we consider all our variables, not just those used to create the clusters. Cluster 1 individuals (N = 61) performed most poorly on all reading subtests, and their verbal IQ was significantly lower than for the two good reader clusters (2 and 4). Their VF asymmetries ranged from average to below average, but their planar asymmetry and corpus callosum measurements were not unusual in any way. This group appears to be low in verbal skills, relative to our college population, but we observed no particular neuroanatomical marker associated with this behavioral profile. Since neither their VF, nor their planar, asymmetry was the lowest for our sample, they do not provide evidence for an association between poor reading and lack of asymmetry (see Leonard & Eckert, 2008 for review of this controversy). Cluster 3 individuals (N = 63), in contrast, were quite average readers, and with the exception of one task (masked word recognition), they had exaggerated RVF/LH advantages on our verbal tasks. Masked word recognition differs from the other tasks in that it places high demands on the ability to rapidly extract visual information necessary for word recognition, and engaging in this process apparently involves more bilateral processing for persons in this cluster. However, although this group had the largest asymmetries for most tasks, their neuroanatomical measurements were quite typical for the sample. In this sample, then, the direction/degree of planar asymmetry did not distinguish among college student readers with average or below average reading skill.

Individuals in Clusters 2 (N = 26) and 4 (N = 33) were highly skilled readers, as evidenced by their scores across all the reading subtests. They differed, however, in their VF asymmetries and in callosal measurements. Cluster 2 was characterized by consistently small VF asymmetries, and interestingly, very large splenia. The splenium area for this group was significantly larger than for all other clusters. These findings are consistent with the idea that reduced task asymmetries are associated with a greater degree of interhemispheric communication. Because the splenium connects visual processing areas, this may explain the association of this callosal region with performance in our divided visual field tests. Some have argued that mixed handers have a greater degree of interhemispheric communication than consistent handers (Christman, Jasper, Vralakshmi, & Cooil, 2007; Christman, Propper, & Dion, 2004). To the extent that callosal area indexes facility of interhemispheric communication, our findings do not support this view as the clusters did not differ by any measure of handedness. The VF and callosal data for Cluster 2 do suggest an association between large splenia and reduced visual, but not manual, asymmetry.

Cluster 4 individuals also had strong reading scores but a more complex pattern of VF asymmetries, characterized by extremely large masked word recognition asymmetry, reduced asymmetry for nonword naming, and average asymmetry for lexical decision. Most VF measures indicated larger RVF/LH advantages for Cluster 4 individuals, as compared to the good readers of Cluster 2. The posterior body of the corpus callosum was smallest for Cluster 4 individuals. It is interesting that differences in (posterior) callosal areas were only observed for the clusters with highly skilled readers, with an association between reduced VF asymmetry and enlarged splenia (Cluster 2) and larger (although variable) VF asymmetry and reduced callosal posterior body (Cluster 4). Prior studies of the relation between callosal area and reading ability have contrasted normal and dyslexic readers (see review in Fine, et al., 2007), but cannot inform us about callosal morphology among highly skilled readers, whereas studies of language lateralization and callosal size have not investigated reading skill effects (e.g., Westerhausen, et al., 2006). The corpus callosum continues to mature well into adulthood (Pujol, Vendrell, Junque, Marti-Vilalta, & Capdevila, 1993), with extended maturation for posterior areas, particularly the splenium, as indicated by age-dependent variations in macrostructure (Giedd, et al., 1999), axonal organization evident from diffusion tensor imaging (Li & Noseworthy, 2002; Muetzel, et al., 2008), and behavior (Muetzel, et al., 2008). The most proficient readers will have accumulated greater reading experience during this later maturation period, perhaps increasing the probability of experience-dependent sculpting of callosal organization relevant to reading processes. One can speculate that these skilled readers continue to “fine tune” the relationship between hemispheric specialization for reading and interhemispheric channels, resulting in associations between lateralization and callosal area that are not found in less skilled readers. However, given the exploratory nature of the current study, this suggestion should be regarded as highly speculative.

In general, the cluster analysis findings suggest that a variety of reading/lateralization profiles exist among college age readers, and that these cannot be accounted for by variations in standard grouping variables such as sex or handedness. In addition, this bottom-up approach revealed associations between reading skill and VF lateralization that could be obscured by correlational approaches. For example, although the clusters differed in reading skill this was not related to VF lateralization for word reading in a linear way: uniformly small VF asymmetries were observed in one group of skilled readers, moderately small or variable asymmetries were observed in both another group of skilled readers and in poorer readers, while the largest VF asymmetries were found to characterize those with average reading skill. In addition, although the clusters were determined based on behavioral data alone, variations in posterior corpus callosum anatomy were observed, implying that clusters of traits related to reading can be associated with particular neural substrates.

Characteristics of Multivariate Outliers

A minority of individuals in our sample had reading and VF asymmetry profiles that differed substantially from the rest of the sample. This suggests that their performance departed in idiosyncratic ways from the “normative” profiles identified by the cluster analysis. In an effort to understand what made these individuals unusual we compared their behavioral and neuroanatomical findings to the rest of the sample. Some intriguing results were observed. First, the outliers had significantly higher passage comprehension scores than the clustered individuals, yet their word level reading ability was unremarkable. This contrasts to those in clusters 1–4 whose word- and text-level reading skills were quite similar (see Tables 1 and 2). This could indicate that the outliers achieve superior comprehension via increased reliance on top-down context information that is available for text, but not single word, reading. The higher IQs of the outliers are consistent with this interpretation as greater world knowledge and reasoning abilities could serve to enhance comprehension independent of word identification skill (c.f., Chiarello, Lombardino, Kacinik, Otto, & Leonard, 2006). Second, the outliers tended to have greater RT, but not accuracy, asymmetries than the rest of the sample (see Table 5, Composite VF Asymmetry). This decoupling of RT and accuracy asymmetry was not observed for the combined cluster group, nor for any individual cluster. Third, both accuracy and especially RT asymmetries were very inconsistent from task to task for the individuals identified as outliers relative to the rest of the sample (see Table 5, Consistency of VF Asymmetry). In other words, these individuals had very large discrepancies in the size and direction of their VF asymmetry indices across tasks, for both response measures⁵.

In general, then, the behavioral outcomes for the outliers were characterized by dissociations between measures that tended to covary for the rest of the sample: between word-and text-level reading ability, between RT and accuracy asymmetry, and between VF asymmetries across various lexical tasks. This explains, at least partially, why these individuals were statistically identified as outliers relative to the rest of the sample. Conceptually, these dissociations suggest a less regulated form of behavioral development, at least for abilities related to reading.

A biological framework for such an outcome is provided by the concept of buffering, that is, reduced sensitivity of phenotypes to genetic and environmental influences during development (Salazar-Ciudad, 2007). This view is based on current research in developmental biology suggesting that a complex genetic network regulates developmental processes that serve to buffer the organism from random influences (Siegal & Bergman, 2002; Rice, 2008). The net effect of this regulation is “canalization”, that is, silencing of genetic variation and regression toward the population mean as originally suggested many decades ago (Waddington, 1957). In other words, individuals with a greater degree of buffering will have more “typical” phenotypes. This more contemporary research (e.g., Salathia & Queitsch, 2007; Landry, 2009) provides a molecular basis for the earlier concept of developmental instability (Markow, 1994; Yeo, et al., 1993). It is important to note that individuals will differ in the effectiveness of regulation (Jaenisch & Bird, 2003; Rasmuson, 2002). We hypothesize that the “outliers” identified in our study may have a less buffered type of development that is more open to the effects of random genetic and environmental perturbations. Random genetic/environmental influences could have both positive and negative effects on behavioral outcomes in individuals with less regulated, more plastic development (Belsky & Pluess, 2009). In our sample, the outliers generally had superior intelligence and reading comprehension. This may be due to the fact that our sample was drawn from a university community. In contrast, dysregulated individuals who had been subjected to primarily negative environmental and genetic influences might be much less likely to develop the cognitive and linguistic abilities needed to attend college, and hence not be represented in our sample population. Clearly, a sample that is more representative of the general population will be needed to examine this conjecture.

The pattern of planar asymmetries observed for the outliers is consistent with the dysregulation view. Within this framework, development of the right and left hemisphere would be coordinated in individuals with high degrees of epigenetic regulation, resulting in modal asymmetries. Relative to the rest of the sample, the outliers were more likely to have reversed (rightward) planum temporale asymmetry (Figure 1), with the reduction of asymmetry due to increased size of the typically smaller structure in the right hemisphere. It is notable that asymmetry for the planum temporale is typically very robust, having been replicated in numerous studies (Shapleske, Rossell, Woodruff, & David, 1999; Sommer, et al., 2008). Because the outliers were identified as unusual based only on their behavioral findings, it is quite interesting that they were also unusual in neuroanatomy.

Finally, we note that findings from the outliers provide counterevidence for the view that dyslexia or impaired reading is associated reduced or reversed planar asymmetry (Leonard & Eckert, 2008). This group did not show the expected leftward planar asymmetry, yet their reading comprehension was superior. Hence the current data imply that poor reading is not a necessary correlate of unusual planar asymmetries.

Strengths and Limitations

An important strength of the current investigation is the large sample size that allowed us to examine a range of behavioral and neuroanatomical variation within a population of normal readers. We were also able to assess groupings based on clusters of behavioral traits instead of relying on a priori groups based on a single behavioral measure (Golestani, et al., 20078; Prat, et al., 2007a). Furthermore, this method allowed us to investigate individuals who did not fit into any behaviorally defined group (i.e., outliers), rather than having such persons contribute to “error variance.” To the extent that our bottom-up approach produced meaningful behaviorally-defined groups, we could begin to investigate dimensions of individual difference that have not emerged using more traditional methodological approaches.

However, we note that our sample, although large, cannot represent the extent of variation present in population as a whole. Because all participants were college students, individuals with the lowest levels of ability, achievement, or socioeconomic opportunity were not adequately represented. As noted above, this prevented us from investigating potentially dysregulated individuals whose environmental/genetic influences might have resulted in negative outcomes. A much broader, community-based sample would be needed to fully exploit the range of variation present in the population as a whole. Yet, even with our restricted sample we observed extensive behavioral and neuroanatomical variability. The small number of left-handers in the sample also limits our ability to address issues regarding the direction (as opposed to degree) of handedness. Additional research that over-samples the left-handed population will be required to investigate structural/functional correlates of handedness direction.

Another limitation concerns our methods for assessing brain structure and function. We relied on corpus callosum area measurements that cannot reveal variations in axonal integrity or connectivity. Investigations using diffusion tensor imaging are needed to more fully explore individual differences in callosal connectivity. The planum temporale measure relied on sulcal boundaries that do not necessarily indicate cytoarchitectural borders (Fischl, et al., 2008), although these measurements have been shown to predict behavior in a number of studies (e.g., Eckert, et al., 2008; Leonard, et al., 1996). Finally, the VF method is a very indirect way to assess functional lateralization, although it does permit assessment of multiple tasks from a large sample. Ideally what is needed is functional imaging performed in a way that takes into account individual variation in brain structure (e.g., Crosson, et al., 1999; Devlin & Poldrack, 2007), so that variations in behavior, structure, and regional brain activity can be examined concurrently.

Characterizing Individual Differences in Brain-Behavior Relations

Interest in individual differences often stems from a desire to understand what makes each of us unique. Yet, paradoxically, we must scientifically explore this issue by exploiting similarities among persons, either by treating each person as a single point on a continuum of variation (regression approach) or by combining individuals that are similar in some way into subgroups. The cluster analysis method employed here is simply a way to identify subgroups that makes fewer assumptions about the dimensions along which individuals should differ. The data we report suggests that this approach can reveal some interesting variations in reading skill, lateralization, and brain structure that might not be evident using other methods. However, it is important not to reify the particular subgroups identified by this analysis. Had we entered different measures into the cluster analysis, different subdivisions within our sample would no doubt have been observed. Because each individual differs from others on a nearly infinite variety of dimensions, there will be many, many ways in which this multidimensional space can be partitioned. We make no claim that the way we partitioned this space in the current investigation is necessarily more valid than another.

To illustrate this point, consider the relationship between reading ability and lateralization (structural and functional). A continuing question for neuropsychological investigations of reading is whether or not reduced or reversed cerebral asymmetry represents a risk factor for reading acquisition (Leonard & Eckert, 2008), and whether variations in adult reading skill covary with lateralization differences (Chiarello, et al., 2009a). In a previous publication on the current sample, we reported significant positive correlations between reading skill and VF lexical lateralization, but only for those with consistent handedness (Chiarello, et al., 2009a). The correlations for consistent handers were quite modest (approximately .24), but suggested that for those with strongly expressed handedness, better reading was associated with larger RVF/left hemisphere advantages. Yet in the current paper using the same sample, we report four subgroups characterized by differing reading/VF lateralization profiles. The groups did not differ by any measure of handedness, and the cluster with the largest RVF advantages had average (for our sample) reading skill. The most skilled readers (Clusters 2 and 4) did not have the greatest VF asymmetries. Because both sets of findings were obtained from the same sample, we cannot attribute the differing results to different experimental methods or demographic characteristics. Rather, we think that we have observed different ways of carving up the same multidimensional individual difference space. Note that even for consistent handers only approximately 6% of the variance in reading was accounted for by variations in VF lateralization (Chiarello, et al., 2009a). At least some of this unexplained variance is likely due to the group differences uncovered by the current cluster analysis.

A visual metaphor may help to clarify this situation. When looking through a kaleidoscope, we can perceive a succession of different visual patterns by rotating the kaleidoscope tube that adjusts a set of internal mirrors. The elements producing these patterns (colored beads) do not change as the tube is rotated, and none of the resulting patterns is a truer reflection of the elements than any other. Yet a multitude of different patterns can be observed. Similarly, as investigators we strive to find patterns in the variations we observe between individuals. The patterns we can observe depend upon the settings we have selected for our analytical lens (i.e., the variables and statistical methods we select). Only by continually rotating the tube (i.e., varying our variables and methods) can we hope to understand the many ways in which individuals are similar and different. It is likely that by so doing we will uncover a variety of ways in which the human brain can support cognitive functions such as reading.